Deep Learning for Document Layout Generation: A First Reproducible Quantitative Evaluation and a Baseline Model

Deep generative models have been recently experimented in automated document layout generation, which led to significant qualitative results, assessed through user studies and displayed visuals. However, no reproducible quantitative evaluation has been settled in these works, which prevents scientific comparison of upcoming models with previous models. In this context, we propose a fully reproducible evaluation method and an original and efficient baseline model. Our evaluation protocol is meticulously defined in this work, and backed with an open source code available on this link: https://github.com/romain-rsr/quant_eval_for_document_layout_generation/tree/master.